Forge-Agentic-Coding-CLI

About

Forge was built as an alternative to trusting a hosted agent with your repository: everything runs locally with no telemetry, actions pass through a default-deny permission system with a realpath-confined filesystem and destructive-command blocking, and secrets live in the OS keychain. Requests flow through a plan-first loop — classify, plan into a DAG, user approval, bounded execution, validation gate, reviewer — with retry capped at three before a debugger diagnosis phase. Local model runtimes (Ollama, LM Studio, vLLM, llama.cpp) are auto-detected and prompts are adapted across 41 classified model families per agent role, with hosted Anthropic/OpenAI-compatible access optional. Nine enforceable modes with per-mode budgets, a local web dashboard, a GitHub Action, and 24 subcommands round out a runtime aimed at developers who want Claude Code-style capability without cloud dependency.